DBSCAN

DBSCAN is a popular algorithm for finding clusters of spatial data. It identifies core points that have enough (defined by the user) neighbors within some distance (also user defined). Points that are not core points but are within the distance of a core point are considered border points of the cluster. Points that are not core points and are not within the distance of a core point are considered outliers and not part of any cluster.

The algorithm requires two parameters:

• epsilon – The farthest apart two points can be while still being considered connected or related. epsilon must be a positive double float.
• minPoints – The minimum number of neighbor points (as determined by epsilon). A point needs minPoints neighbors to be considered a core point. minPoints must be a positive integer.

Example overview

In this example, we will generate some random data and use DBSCAN to cluster that data. Then, we’ll visualize the clusters using a scatter plot.

This demo is derived from the scikit-learn DBSCAN demo.

%pip install scikit-learn

Define Sedona Context

from sedona.spark import *

config = SedonaContext.builder().getOrCreate()
sedona = SedonaContext.create(config)

Data Generation

In the following code section, we’ll generate some data using sklearn’s make_blobs function. We’ve set the data to consist of 750 points with 3 clusters. After clustering the data, we’llvisualize it in pyplot.

from sklearn.datasets import make_blobs
from sklearn.preprocessing import StandardScaler
import matplotlib.pyplot as plt

center_clusters = [[1, 1], [-1, -1], [1, -1]]
feature_matrix, labels_true = make_blobs(
   n_samples=750, centers=center_clusters, cluster_std=0.4, random_state=0
)

feature_matrix = StandardScaler().fit_transform(feature_matrix)

plt.scatter(feature_matrix[:, 0], feature_matrix[:, 1])
plt.show()

Clustering

In the following section, we’ll use the DBSCAN implementation in Wherobots to cluster the data in a dataframe, setting epsilon to 0.3 and minPoints to 10.

Wherobots’ DBSCAN returns outliers by default.

from sedona.spark import *

df = sedona.createDataFrame(feature_matrix).select(ST_MakePoint("_1", "_2").alias("geometry"))
clusters_df = dbscan(df, 0.3, 10, include_outliers=True)

clusters_df.show()

Visualization

Finally, we’ll visualize the clusters with matplotlib. ST_X and ST_Y pull the point coordinates out of the geometry column, and some manipulations are made to the data to improve the clarity of the visualization.

import matplotlib.pyplot as plt

pdf = (clusters_df
       .selectExpr("ST_X(geometry) AS x", "ST_Y(geometry) AS y", "cluster", "isCore")
       .toPandas()
      )

fig, ax = plt.subplots(figsize=(10, 8))
ax.scatter(
    pdf["x"], pdf["y"],
    # One color per cluster (outliers get their own)
    c=pdf["cluster"].astype("category").cat.codes, cmap="tab20",
    # Core points are drawn twice as large as border points and outliers
    s=(pdf["isCore"].astype(int) + 1) * 40,
    edgecolors="lightgray",
)
ax.set_title("DBSCAN clusters")
plt.show()